Kurt Smith wrote:
> Hi All,
>
> It was suggested off list that I gather some simple benchmarks
> comparing the different ways of passing a struct with a contiguous
> array pointer inside -- through a copy, through a pointer, and through
> explicit unpacking of the struct.
>
> So far, I find almost no difference whatsoever between the 3 modes,
> which is encouraging. If these benchmarks hold up, it means we can
> choose the 'best' way to pass Py_buffer structs around without worry
> of a speed hit.
>
> Comments welcome.
Nice. I do have some comments to refine this:
- Note that your "simp" vs. "unpack" are likely identical in
assembler, so you don't need to have them both I thinkk.
- We also want to compare Py_buffer vs. simpler formats. E.g. if you
want to pass a simple contiguous buffer, do you pass Py_buffer or your
"simp"?
- Py_buffer is much larger than your simp, which could affect the
benchmark for ptr vs. non-ptr.
- Make sure ll is Py_ssize_t, this could affect things on 64-bit systems.
- In Py_buffer, the shape is stored as a pointer, so you actually need
an extra pointer dereference to get the shape: buf.shape[0] vs. simp.ll.
So the next step would be actually using the Py_buffer struct (which is
a Cython builtin and is documented in PEP 3118), and see how Py_buffer
(either by copy or by ptr) holds up against simp.
As Py_buffer would be filled out outside of what you do timings on, you
don't need to fill in things besides what you need, this will suffice:
cdef Py_buffer buf
buf.shape = &sp.ll # array of lenght 1 :-)
# Will cause big problems if sp.ll is not Py_ssize_t!
buf.buf = sp.dta
--
Dag Sverre
_______________________________________________
Cython-dev mailing list
[email protected]
http://codespeak.net/mailman/listinfo/cython-dev
